Kubuxu

Microchip also wants money for license to their design suite.

So you read an article "China disappears random person" and then at the end you click a button and bet whether that person is alive or dead or whatever.

Congratulations, you just created a murder market. Create “Is person X dead?”, and put a large limit order buying No. Anyone with means can capitalise on it by fulfilling the physical obligation of the market and buying up the limit order. Don’t create the murder market.

---

It would be a stupid thing to do given the trail, bit it gives the person ordering it sort of plausible deniability.

I have TDD flow working as a part of my tasks structuring and then task completion. There are separate tasks for making the tests and for implementing. The agent which implements is told to pick up only the first available task, which will be “write tests task”, it reliably does so. I just needed to add how it should mark tests as skipped because it’s been conflicting with quality gates.

That somewhat existed in servers in the past, my R720xd had RAM mirroring mode. IDK if it used it for reducing latency, but you could take out a stick and the server would continue running as normal and report an alarm in iDRAC.

Yeah, after two paragraphs, my eyes started skimming it due to low signal-to-noise ratio.

Just try using Claude with API for an hour and you will see that the subscriptions are definitely not profitable (unless they percent off “partying but dormant” is very high).

Reasoning, in majority of cases, is pruned at each conversation turn.

A paper on the same topic: On the Expressiveness of Softmax Attention: A Recurrent Neural Network Perspective, Gabriel Mongaras, Eric C. Larson, https://arxiv.org/abs/2507.23632

Video presentation if someone prefers it: https://www.youtube.com/watch?v=PN3nYBowSvM

Linear attention is a first-degree approximation of Softmax attention, and model performance gets better as you increase the degree of the Taylor approximation.

I'm thinking about adapting an existing model to Taylor-approximated attention. I think it should be possible with some model surgery and rehabilitation training.

Definitely agreed, that is what I do as well. At that point you have good understanding of that code, which is in contrast to what the post I responded suggests.

I don’t see how I would feel comfortable pushing the current output of LLMs into high-stakes production (think SLAs, SRE).

Understanding of the code in these situation is more important than the code/feature existing.

And DEF is ~33% urea in water.

I've done that in the past (8+ years ago) with raw IP sockets.

For 8+ bays you just need a SAS HBA card and one free PCI-E slot. Not to mention that many motherboards will have 6+ SATA ports already.

If anything, 2nd hand AMD gaming rigs make more sense than old servers. I say that as someone with always off r720xd at home due to noise and heat. It was fun when I bought it during winter years ago, until summer came.

Traditional CMOS D-flipflop takes 16 transistors, SRAM cell takes 6 (two looped inverters and two data-line selection transistors).

The capacitor in DRAM is usually realised as an enlarged gate of a transistor AFAIK.

It is always PWM under the hood, the question is, how much was spent (or not) on the filtering network out of the PWM. Is it closer to buck converter or is it straight up flicker at the output.

Nobody runs 4096 samples per pixel. In many cases 100-200 (or even less with denoising) are enough. You might run up to low-1000 if you want to resolve caustics.

IIRC reordering additions can cause the result to change which makes auto-vectorisation tricky.

You can think of closing the channel as sending a message “there will be no further messages”, the panic on write is enforcement of that contract.

Additionally the safe way to use closing of a channel is the writer closing it. If you have multiple writers, you have to either synchronise them, or don’t close the channel.

Majority of good spots for hydro were already built up, and if they weren’t, good luck with NIMBY.

Beam Search sampling is sometimes getting used

The question is, "If I look at this repo, who owns the copyright?" Sure, you could move the original LICENSE into a directory. Still, if the files are intermingled, you should have a prominent notice that says, " Hey, these files have mixed copyright ownership."

In case of deps, the dependency comes with its own LICENSE file.

In this case the code is essentially forked, integrated and intermingled, so that is why it should be in the LICENSE file.

If it was file or two, it would be fine to add a comment pointing to the license file in the repo, if it was a directory, or to copy it verbatim to that file. It all the copied code was in a directory then having it in directory would be fine.

In this case it looks like they took the original code and heavily modified it, so the simplest way to solve it is one LICENSE with both notices.

The question is who does the copyright belong to in this repository. It is both original author and Microsoft (because they took authors code and modified it). So the License file should mention both.

TME is European as well, and I think Arrow has warehouses in Europe.

Could you list few? I’ve stopped following media in general because everything I could find was profit maximisation through outrage and sensationalism. So I would really appreciate something that isn’t that.

I think what would be useful is a histogram of population densities, or for example a percentile based metric: population density covering 99/95% of population.

Don’t be sad for commercial entity that is not a good player https://github.com/ggerganov/llama.cpp/pull/11016#issuecomme...

I don't think entire GPU is specialised nor a singular token will use the same expert. I think about it as a gather-scatter operation at each layer.

Let's say you have an inference batch of 128 chats, at layer `i` you take the hidden states, compute their routing, scatter them along with the KV for those layers among GPUs (each one handling different experts), the attention and FF happens on these GPUs (as model params are there) and they get gathered again.

You might be able to avoid the gather by performing the routing on each of the GPUs, but I'm generally guessing here.

Not sure about DeepSeek R1, but you are right in regards to previous MoE architectures.

It doesn’t reduce memory usage, as each subsequent token might require different expert buy it reduces per token compute/bandwidth usage. If you place experts in different GPUs, and run batched inference you would see these benefits.

I remember reading a paper that showed that giving models even a a few filler tokens before requiring a single phrase/word/number answer significantly increasee accuracy. This is probably similar.